Extracting the Essence from Sets of Images

Abstract

We use a set of photographs showing similar scenes as a model for a single photograph this scene. A distance measure for this model is defined by correlating the neigborhoods of pixels in similar positions. A cross analysis of the source images yields confidence values for their pixels. The confidence values together with the distances in pixels are used to steer a variable bandwidth mean shift algorithm that moves an arbitrary image towards one conforming with the model. Furthermore, distances are also used for a non-local means reconstruction of image areas that have no consistent explanation in the source images. This allows reconstructing images of scenes that are inconsistently documented in the source images, e.g. are occluded in the majority of images.